denoise signal representations

The necessity of noise conditioning in denoising generative models remains unchallenged despite its ubiquitous adoption. Method: We systematically evaluate the impact of removing noise conditioning across diverse denoising architectures via theoretical error analysis, ablation studies, and FID-optimized unconditional sampling. Contribution/Results: Contrary to prevailing assumptions, most denoising models exhibit robust performance without noise conditioning—and in several cases, achieve lower FID scores than their conditioned counterparts. We introduce the first high-performance noise-unconditional diffusion model, attaining a FID of 2.23 on CIFAR-10—narrowing the gap with state-of-the-art conditional models significantly. Our findings demonstrate that the denoising generative paradigm need not rely on explicit noise conditioning, opening new avenues for architectural simplification, computational efficiency gains, and foundational theoretical reexamination.

To address the degradation in generation quality caused by misalignment between noise levels and distances to the data manifold during diffusion model denoising, this paper proposes a noise-level calibration mechanism. It explicitly models noise level as a proxy for the distance from a sample to the data manifold and introduces a lightweight, plug-and-play auxiliary correction network that dynamically refines denoising estimates at each step. The method requires no retraining of the backbone diffusion model and uniformly supports diverse restoration tasks—including image inpainting, super-resolution, deblurring, color enhancement, and compression artifact removal—while remaining compatible with mainstream samplers (e.g., DDIM). Correction subnetworks are constructed solely from pre-trained denoising networks, incorporating manifold geometric priors and task-specific constraints (e.g., masks, degradation kernels, frequency-domain restrictions). Experiments demonstrate consistent improvements in FID and LPIPS across both unconditional generation and restoration tasks, with minimal computational overhead and stable performance gains.

Real-world image denoising faces dual challenges: poor generalizability of handcrafted priors and the heavy reliance of deep learning methods on large-scale paired noisy-clean training data. To address this, we propose Net2Net—a novel framework that, for the first time, seamlessly integrates unsupervised Deep Image Prior (DIP) with a supervised pre-trained denoiser (DRUNet) under a unified Denoising-based Regularization (RED) optimization scheme, requiring no paired annotations. Net2Net synergistically leverages the input-specific modeling capability of untrained networks and the rich noise statistics encoded in large-scale pre-trained models, achieving strong generalization across diverse noise types and imaging conditions without compromising inference efficiency. Extensive experiments on multiple real-world denoising benchmarks demonstrate that Net2Net significantly outperforms existing state-of-the-art methods—especially under extreme data scarcity—while maintaining lightweight deployment.

Existing learning-based image denoising methods rely on fixed noise priors, leading to poor generalization under varying real-world noise distributions. To address this, we propose a novel paradigm that decouples noise priors from image priors, and introduce a conditional optimization framework capable of estimating sensor-level noise priors directly from a single sRGB noisy image. Our key contributions are: (1) the first explicit modeling and incorporation of noise priors into denoising architectures; (2) a lightweight Local Noise Prior Estimator (LoNPE) network for pixel-wise noise prior estimation; and (3) a Conditional Denoising Transformer (CondFormer) that dynamically injects estimated noise priors into the denoising subspace via conditional self-attention. Extensive experiments demonstrate significant improvements over state-of-the-art methods on both synthetic and real-world datasets, with strong cross-device robustness and generalization capability. The code is publicly available.

This work addresses high-fidelity image reconstruction from degraded observations. Methodologically, it proposes an adaptive diffusion framework that integrates conditional sampling with pre-trained diffusion priors (e.g., Stable Diffusion) and introduces a novel vision-language model (VLM)-guided nearest-neighbor image retrieval strategy to dynamically align the diffusion process with input degradation characteristics—overcoming the limitations of fixed priors. A lightweight fine-tuning mechanism enables end-to-end adaptation. Experiments demonstrate state-of-the-art performance on super-resolution, motion deblurring, and text-driven image editing, significantly outperforming existing approaches. The results validate the effectiveness of the adaptive diffusion paradigm in harmonizing observational consistency with natural image priors, while exhibiting strong generalization across diverse restoration tasks.

This work investigates the mechanisms by which diffusion models generate highly realistic images that differ from their training data—referred to as “creativity”—and demonstrates that this capability stems from the alignment between the denoiser architecture and the target data distribution. Through theoretical analysis and empirical experiments, the study derives explicit forms of the generated distribution for linear, polynomial, and bottleneck-style denoisers for the first time, and systematically evaluates the behavior of various architectures, including UNet variants, throughout the diffusion process. The findings reveal that minor architectural modifications to the UNet significantly impact generation fidelity, thereby underscoring the critical role of the denoiser’s inductive bias and its alignment with the target distribution in determining model performance.

This work addresses the limitations of existing deep learning approaches for RAW image denoising, which often neglect classical denoising priors, resulting in overly complex models with limited generalization. To overcome this, we propose the first learnable non-local module that explicitly embeds the classical non-local self-similarity prior into a neural network. Our method integrates multi-scale feature extraction, learnable matching and collaborative filtering, noise-level map conditioning, and joint training on both synthetic and real-world noise. The resulting model achieves performance comparable to state-of-the-art CNN- and Transformer-based methods across multiple benchmarks and real datasets, while significantly reducing parameter count and demonstrating strong cross-sensor generalization capability.

This work addresses the challenge of signal-dependent Rician noise in accelerated diffusion-weighted imaging (DWI), which hinders effective denoising by conventional convolutional methods. To overcome this limitation, the authors propose a noise-aware, attention-driven denoising framework that innovatively integrates explicit noise-level conditional embeddings, hierarchical Swin Transformer window attention, and a Transformer-based multidimensional gated refinement mechanism. The architecture further incorporates residual learning and channel-adaptive attention to enable adaptive modeling of heteroscedastic noise. Evaluated across noise levels ranging from 1% to 15%, the method achieves an average PSNR of 33.69 dB and SSIM of 0.8539, demonstrating exceptional robustness and generalization even under severe noise conditions.

This work addresses the challenge of efficiently adapting the denoising dynamics of a pre-trained diffusion model for discriminative representation learning while preserving its generative capabilities. The proposed approach treats noisy latent variables at different denoising timesteps as multi-view augmentations of the same image. By freezing the Stable Diffusion backbone and applying lightweight fine-tuning via LoRA, the method introduces, for the first time, a joint optimization framework combining noise-level contrastive learning with reconstruction loss. This enables synergistic learning between generative and discriminative objectives without requiring training from scratch. Experimental results demonstrate strong performance: a linear probe achieves 80.1% top-1 accuracy on ImageNet-1K, and unconditional image generation at 256×256 resolution yields an FID of 5.56.

This work addresses the challenge of high-fidelity restoration of images corrupted by additive white Gaussian noise (AWGN) at a high noise level (σ=50). Organizing and evaluating advanced deep denoising models submitted by 20 top-performing teams, the study pursues optimal performance in terms of peak signal-to-noise ratio (PSNR) without imposing constraints on model size or computational cost. The final solutions, selected from 116 registered participants, integrate a diverse array of neural network architectures, collectively reflecting the state-of-the-art in unconstrained image denoising. This effort establishes a new performance benchmark for the task, offering a comprehensive snapshot of current methodological advances in the field.